promotion/Llama-3.1-8B-FixedReferenceNBPO

TEXT GENERATIONPricing:Input $0.2 / Cached $0.028 / Output $0.32Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 5, 2026License:llama3.1Architecture:Transformer Featherless Exclusive Cold

Llama-3.1-8B-FixedReferenceNBPO is an 8 billion parameter model based on Meta's Llama-3.1-8B-Instruct, fine-tuned using a Fixed Reference Nash Bargaining Preference Optimization (NBPO) approach. It is specifically optimized for instruction following, truthfulness, honesty, and helpfulness, utilizing the UltraFeedback panel for objective weighting. This model serves as a direct comparison point for NBPO methods, demonstrating performance against a common reference policy.

Loading preview...

Model Overview

promotion/Llama-3.1-8B-FixedReferenceNBPO is an 8 billion parameter language model derived from meta-llama/Llama-3.1-8B-Instruct. This model was developed as part of research into Nash Bargaining Preference Optimization (NBPO), specifically to provide a fixed-reference control for comparative evaluation.

Key Characteristics

  • Base Model: Built upon meta-llama/Llama-3.1-8B-Instruct.
  • Optimization Method: Utilizes a fixed-reference approach within the Nash Bargaining Preference Optimization framework, ensuring the same response pool, temperatures, disagreement point, and optimizer budget as NBPO, but with a different objective weighting strategy.
  • Training Objectives: Focused on improving performance across four key areas:
    • Instruction following
    • Truthfulness
    • Honesty
    • Helpfulness
  • Training Data: Fine-tuned using the UltraFeedback panel.
  • Training Budget: Underwent 300 optimizer updates with a global batch size of 16.

Evaluation and Purpose

This model's primary purpose is to serve as a controlled baseline for evaluating the effectiveness of NBPO. It is evaluated by measuring its independent objective-wise win rate against the common reference policy, with judgments provided by Llama-3.3-70B-Instruct on held-out, prompt-disjoint data. This allows for direct comparison of how different objective weighting strategies impact model performance on critical alignment metrics.